water
Article
A Graph-Based Optimization Framework for Large Water
Distribution Networks
Robert Sitzenfrei
Citation: Sitzenfrei, R. A Graph-
Based Optimization Framework for
Large Water Distribution Networks.
Water 2023, 15, 2896. https://
doi.org/10.3390/w15162896
Academic Editor: Pankaj Kumar
Received: 11 July 2023
Revised: 8 August 2023
Accepted: 9 August 2023
Published: 11 August 2023
Copyright: © 2023 by the author.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Water 2023, 15, 2896. https://doi.org/10.3390/w15162896
Unit of Environmental Engineering, Department of Infrastructure Engineering, University of Innsbruck,
6020 Innsbruck, Austria; robert.sitzenfrei@uibk.ac.at
Abstract: Water distribution networks (WDNs) have a crucial task: to reliably provide sufficient and
high-quality water while optimizing financial resources. Achieving both reliability and resilience
is vital. However, oversizing capacities can be costly and detrimental to water quality due to
stagnation. Designing WDNs requires the consideration of these factors, resulting in a multi-objective
optimization task typically addressed with evolutionary algorithms. Yet, for large WDNs with
numerous decision variables, such algorithms become impractical. Complex network analysis offers
an efficient approach, particularly with mathematical graphs representing WDNs. Recently, a graph-
based multi-objective design approach using a customized measure (demand edge betweenness
centrality) and a surrogate method for water quality assessment in large WDNs were developed. This
paper combines these graph-based approaches into an optimization framework suitable for complex,
real-world WDNs. The framework aims to minimize costs, maximize resilience, and exclude designs
with poor water quality. It is demonstrated on a toy example, and its computational efficiency is
shown by a real case study with 4000 decision variables, obtaining results in just 18.5 s compared to
weeks of computation time with a state-of-the-art evolutionary algorithm.
Keywords: multi-objective optimization; demand edge betweenness centrality; hydraulically
informed graph analysis; water quality surrogate model; least cost design
1. Introduction
The major task of water distribution networks (WDNs) is to reliably supply water in
sufficient quantity and quality with the most efficient utilization of financial funds (i.e.,
minimum costs) at the same time [1]. While the reliability of a WDN can be interpreted as
the probability that it operates in a satisfactory state under normal conditions, the resilience
of a WDN indicates the system reserve under unexpected and critical conditions [2]. The
oversizing of capacities favors resilience, but at the same time, it is costly, and the water
quality performance due to stagnation is negatively impacted [3].
Further, such systems are usually grown over decades and have an organic and
complex structure [4]. The complex hydraulic interactions of different components is no
trivial problem to solve.
Due to the complexity in design and operation of WDNs, and to ensure a reliable
and resilient level of service with minimum costs at the same time, multi-objective design
approaches are used which are usually solved with evolutionary algorithms [5]. These
approaches are well investigated and described in the literature [6,7]. However, research
on real all-pipe models is rare in this regard due to the large decision space, as for large
WDNs the decision space increases exponentially. When considering multiple objectives
(e.g., resilience, costs, and water quality) for complex, large (real) WDNs with several
thousand decision variables, evolutionary algorithms are practically infeasible to apply
due to the limitations in computation time. However, due to data availability (geographic
information system and digital twins), the trend in modeling is going towards all-pipe
models. Therefore, there is a need for fast approximations for solving such multi-objective
design problems.
https://www.mdpi.com/journal/water
Water 2023, 15, 2896 2 of 15

The topology of WDNs can be modeled as mathematical graphs. With complex net-
work analysis (CNA), such mathematical graphs can be analyzed very computationally
for efficiency, and therefore, such approaches are especially suitable for analyzing large
spatial transport networks [8–13]. Recently, based on CNA, a highly efficient approach for
the Pareto-optimal design (i.e., minimizing costs and maximizing resilience) of WDNs was
developed [14]. Based on topological features and a customized graph measure for the
demand distribution (demand edge betweenness centrality), a graph-based multi-objective
design approach was developed, which outperformed the results of an evolutionary algo-
rithm in some areas of the Pareto-front regarding the quality of solutions and especially
computation time (factor 105 faster). That approach is particularly suitable to solve prob-
lems with an extremely large decision space in acceptable time [14]. However, when
comparing the design solutions from the graph-based optimization with those from evo-
lutionary algorithm, for design solutions with a high level of resilience, the graph-based
method is outperformed. However, in that region of the Pareto-front (high resilience and
high costs), water quality issues arise due to the low flow velocities. But water quality
analysis is even more computationally intensive for large-scale WDN. Recently, based on
CNA, a highly efficient surrogate method for assessing water quality in large WDNs was
developed (2.4 × 105 times faster than extended period simulation Epanet2) [15]. Therein,
the edges in the graph of the WDS are weighted based on the residence time in the pipes.
Based on shortest path analysis, pattern correction, and topological correction functions
(network dispersion), good estimates for nodal water age values were obtained.
In this paper, the approaches based on CNA, the (1) two-objective optimization
model [14] including a hybrid approach for considering topography in graph designs [16]
and (2) the graph-based water quality model [15], are combined in a novel graph frame-
work for the optimization of WDNs, which is especially suitable for complex, large (real)
WDNs. Pareto-optimal design solutions (minimum costs versus maximum resilience) are
evaluated regarding the exceedance of a threshold for water age. Therewith, technically
unsuitable solutions are excluded from the Pareto-front. For a toy example, the three differ-
ent approaches are presented in detail, including the applicability of combining these very
computationally efficient, novel approaches. The applicability of this very computationally
efficient, novel approach is shown on real case studies with 4000 decision variables for
which the results can be obtained within only few seconds of computation time.

2. Materials and Methods

In this work, a graph-based approach is used (Section 2.1) for optimizing WDNs with
two contradicting objectives: minimal costs versus maximum resilience. The approach
involves identifying a Pareto-front of optimal design solutions, solely considering network
topology and demand distribution, which allows for a trade-off between the two objec-
tives. However, especially for least-cost designs, the topography (i.e., the height difference
between sources and sinks) plays an important role. Therefore, a hybrid approach is im-
plemented (Section 2.2), which identifies pressure deficits based on hydraulic simulations
and modifies the graph measures in order to obtain designs that fulfill minimal pressure
requirements. Subsequently, using a graph-based water quality approach (Section 2.3),
design solutions are identified that exceed a water quality threshold, specifically the maxi-
mum water age. This approach allows the evaluation of the quality of the design solutions
obtained from the Pareto-front and selects those that meet the water quality requirements.
The graph-based procedures are step-by-step shown on a small toy example, and the
applicability to a large real case study is shown. Finally, the obtained solutions for the large
case study are compared with design solution based on an evolutionary algorithm [17]
for which extended period simulations with Epanet2 for water quality assessment are
performed to exclude infeasible solutions from a water quality perspective.
 2.1. Graph-Based Multi-Objective Design with Demand Edge Betweenness Centrality
A WDN can be modeled as mathematical graph, which can be analyzed computa-
tionally for efficiency with CNA. A graph consists of a set of vertices (e.g., nodes), which
can be interlinked with links (e.g., pipes). The adjacency matrix, which is a symmetric
Water 2023, 15, 2896
matrix of the number of nodes, describes if there is a link between two vertices. If there is
3 of 15
a connection between vertices i and j, the matrix element aij = 1, otherwise aij = 0. Each
link/edge k in the graph can have a weight wk. Often unweighted graphs (wk = 1) or the
Euclidean distance (i.e., pipe length Lk) is used. But also, other weights can be used, e.g.,
2.1. Graph-Based
mimicking Multi-Objective
hydraulic or water qualityDesign behavior,
with Demand Edge
such as Betweenness
friction losses Centrality
or residence time
in anAedge.
WDN can be modeled as mathematical graph, which can be analyzed computa-
tionally
Thefor efficiency
shortest path,with CNA.
σi,j, is the pathA graph
between consists
two of a set ofi and
vertices, vertices (e.g.,the
j, where nodes), which
path length
can be
(i.e., theinterlinked
sum of edge with links (e.g.,
weights) pipes). The adjacency matrix, which is a symmetric
is minimal.
matrix Theofedge
the number of nodes,
betweenness describes
centrality (EBC) if there is a link
of a graph between
counts how twooftenvertices.
an edgeIfistherepart
is a connection between vertices i and j, the matrix
of σi,j when connecting all possible node pairs. For a WDN, each element a = 1, otherwise
ij node has to be aijconnected
= 0. Each
link/edge
to k in source
at least one the graph cantherefore
node, have a weightcountingwk . the
Often unweighted
shortest paths of graphs (wk all
σS,j from = 1)the
or de-
the
Euclidean distance (i.e., pipe
mand nodes to the source node S can be length L ) is used. But also, other weights can
k an indicator for required transport capacity [18].be used, e.g.,
mimickinginstead
However, hydraulic or water
of simply qualitythe
counting behavior,
numbersuch as friction
of shortest losses
paths or through
going residenceantime edge,in
an edge.
the nodal demand dj can be used to weigh the EBC counts along σS,j. This then gives an
The of
estimate shortest path, σi,jflows.
ideal/design , is theThis
pathcustomized
between two vertices,
EBC measure i and j, where the
is denoted path length
‘demand edge
(i.e., the sum of edge weights)
betweenness centrality’ (EBC ) [14]. Q is minimal.
The edge
When EBCbetweenness
Q is used forcentrality
the design (EBC)of a of a graph
WDN, onlycounts howlength
the pipe often an
Lk edge is part as
is available of
σi,j when connecting all possible node pairs. For a WDN, each node has to be connected to
edge weight, as no other hydraulic characteristics are known before the design process.
at least one source node, therefore counting the shortest paths of σS,j from all the demand
When we consider all edges, which are part of a shortest path from all demand nodes to
nodes to the source node S can be an indicator for required transport capacity [18]. However,
a source node (all edges with EBCQ > 0), this set of edges is called the shortest path tree,
instead of simply counting the number of shortest paths going through an edge, the nodal
and the remaining elements are the co-tree.
demand dj can be used to weigh the EBC counts along σ . This then gives an estimate ofQ
In Figure 1, for the toy example, EBCQ values are S,j determined. Based on the EBC
ideal/design flows. This customized EBC measure is denoted ‘demand edge betweenness
values, the diameters (DNk) for each pipe (k) can be determined with a continuity equa-
centrality’ (EBCQ ) [14].
tion, a flow velocity
Q (vdesign), and commercially available diameter classes (DNavailable):
When EBC is used for the design of a WDN, only the pipe length Lk is available as
edge weight, as no other hydraulic characteristics are known before the design process.
4 EBC (k)
When we consider all edges, DN which = are part∙ of a shortest∈ DNpath from all demand nodes to(1)a
Q π v
source node (all edges with EBC > 0), this set of edges is called the shortest path tree, and
the remaining elements are the co-tree.
Sitzenfrei, Wang, Kapelan, and Savić [14] showed, that the when considering differ-
In Figure 1, for the toy example, EBCQ values are determined. Based on the EBCQ
ent values for vdesign in the described design process, Pareto-optimal design solutions can
values, the diameters (DNk ) for each pipe (k) can be determined with a continuity equation,
be obtained (minimal pipe costs versus maximum resilience according to [19]), which
a flow velocity (vdesign ), and commercially available diameter classes (DNavailable ):
partly outperforms designs from optimization with an evolutionary algorithm. For the
proposed CNA design procedures itself, noQhydraulic simulations are required, only to
4 EBC (k)
check the pressure threshold DNkafterwards
= · (e.g., minimal∈ DN pressure of 30 m under design (1)
load), and a hydraulic simulation run is needed π vdesign [20]. available

Figure 1. Multi-objective
Figure 1. Multi-objectivedesign
design procedure
procedure based
based on CNA,
on CNA, adapted
adapted from from [14] (licensed
[14] (licensed under CC-
under CC-BY 4.0).
BY 4.0).
Sitzenfrei, Wang, Kapelan, and Savić [14] showed, that the when considering different
values for vdesign in the described design process, Pareto-optimal design solutions can be
obtained (minimal pipe costs versus maximum resilience according to [19]), which partly
outperforms designs from optimization with an evolutionary algorithm. For the proposed
CNA design procedure itself, no hydraulic simulations are required, only to check the
pressure threshold afterwards (e.g., minimal pressure of 30 m under design load), and a
hydraulic simulation run is needed [20].
Water 2023, 15, 2896 4 of 15

Further, comparison of these design solutions to those obtained by evolutionary

algorithms showed that a very relevant part of the Pareto front (around the knee bend)
can be determined with the CNA approach. However, design solutions close to the least
cost design cannot compete with those from evolutionary algorithms. The reason for
that is that for the assessment of the performance of WDNs, there are more important
parameters than just the network topology and the demand distribution. Specifically, the
height differences between the sources and the demand nodes are of utmost importance
when determining the least-cost design of a WDN. However, in classical graph-based
approaches, only network topology is considered and solely CNA-based methods are not
able to consider the topography.

2.2. Hybrid Demand Edge Betweenness Centrality and Least Cost Designs
To overcome the lack of design solutions close to the least-cost design, a hybrid ap-
proach was developed, where low-cost design solutions were evaluated with a hydraulic
solver, and subsequently, the demand edge betweenness centrality distribution was itera-
tively altered for nodes with pressure deficits [16].
For the toy example, in Figure 2, it is shown how to determine hybrid EBCQ . The
initial EBCQ distribution is similar to Figure 1. When now applying the continuity equation
(Equation (1)) with a design velocity (vdesign ) of, e.g., 0.8 m/s, the required diameters can
be determined. When using a set of available discrete pipe diameters (e.g., 76.2, 101.6,
and 152.4 mm) and corresponding costs per one-meter pipe (EUR 8, 11, and 16/m), the
diameters can be determined for all pipes (rounding up to the nearest available discrete
diameter), and the total construction costs can be determined. For different values for
vdesign , the diameters can be determined, starting with the lowest vdesign, it is then increased
stepwise (0.8, 1.0, to 1.2 m/s). When the pressure distribution is checked with Epanet2
(EN2), pressure deficits can be determined (e.g., ≤30 m). If there are no pressure deficits,
the design solution is technically feasible, and the next (higher) design velocity is evaluated.
If there are pressure deficits present in that design, it is technically infeasible, subsequently
these solutions are enhanced to meet the pressure criterion.
To improve the design solutions experiencing pressure deficits, the nodes with pressure
deficits (e.g., node 3 in the toy example) are determined with Epanet2. This is followed by
determining the shortest path to the closest source node (σ3,S1 ). Along that path, a fraction
(f) of the demand of the node with pressure deficit is added to the EBCQ (i.e., 1.2 L/s in
the example with f = 1) to (slightly) increase the flow capacity from the source to that node.
The design procedure is repeated, and the pressures are determined again with the new
(hybrid) EBCQ values. This procedure is repeated until there are no pressure deficits in the
system. In the next step, the design velocity is further increased and, if necessary, pressure
deficits are again addressed by iteratively using the proposed hybrid approach.

2.3. Water Quality Assessment with Complex Network Analysis

For water quality assessment with CNA, the shortest path analysis is used again [15].
This time, as an edge weight, residence times in the edges are considered. The shortest
paths can then be interpreted as minimal residence time of water during the transport
from a source to a demand node (i.e., water age or travel time in the WDN). In order to
avoid additional hydraulic simulations, the hydraulic simulation results from the resilience
assessment (see Section 2.1) with the design load (Qdesign ) are used. To account that, for
water age simulation, the average demand load is decisive (Qavg ), the flow velocities from
the resilience assessment are reduced by a factor (fV ) describing the ration between design
and average demand (Qavg = fv Qdesign ). In principal, there could be changes in the flow
regime with different load cases, but Sitzenfrei [15] showed that for typical WDNs with
that assumption, neglectable errors regarding the flow velocities are obtained.
Water 2023, 15,
Water 2023, 15, 2896
x FOR PEER REVIEW 55 of
of 15
15

Figure 2. Hybrid approach for EBCQ for least-cost designs.

Figure 2. Hybrid approach for EBCQ for least-cost designs.
2.3. Water Quality
In Figure Assessment
3, the concept ofwith Complex
water Network
quality Analysis
assessment with CNA is shown and described
For water
in Equations qualityThe
(2)–(4). assessment
determinedwithvalues
CNA,of the
σi,jshortest
for waterpathageanalysis is usedsimilar
are basically again [15].
to a
This time, as an edge weight, residence times in the edges are considered.
hydraulic snapshot simulation, where only edges in the shortest path tree are considered The shortest
in
paths
the can then
model. be interpreted
To account as minimal
for extended residence time
period simulations of different
(i.e., 24 water duringhourlythe transport
values for a
from a source
diurnal demand topattern
a demand node (i.e.,
for multiple water
days), timeageconsideration
or travel timecan in be
theassumed
WDN). In in aorder to
simple
avoid[15],
way additional hydraulic
and a vector simulations,
for pattern the hydraulic
correction (Cp ) can besimulation
determined results
basedfrom thehourly
on the resili-
ence assessment
pattern (see
multipliers Cp2.1) with
i (see the design
Equation (3)).load (Qdesign) are used. To account that, for water
age simulation, the average demand load is decisive (Qavg), the flow velocities from the
resilience assessment are reducedAby CNA a factor + C0 ·eB the ration between design and
= σi,j ·(fCVP) describing (2)
average demand (Qavg = fv Qdesign). In fV principal, there could be changes in the flow regime
with different loadCp i = Cp
cases, but 1 + p with
i−Sitzenfrei [15] i=0 = 0that
Cpshowed andfor ∈ [0; 24WDNs
pi typical ] (3)
with that as-
i
sumption, neglectable errors regarding the flow velocities are obtained.
In Figure 3, the concept B =ofmD
water
2
·Qaltquality
·mixL; assessment
C0 = [0.0; 0.1 with
] CNA is shown and de- (4)
scribed in Equations (2)–(4). The determined values of σi,j for water age are basically sim-
ilar to a hydraulic snapshot simulation, where only edges in the shortest path tree are
considered in the model. To account for extended period simulations (i.e., 24 different
 network correction, the fraction of flow in the co-tree (alternative flow, 0 ≤ Qalt ≤ 1) and the
mean node degree (mD) of the WDN were identified to describe the global network dis-
persion [15]. Qalt is the sum of flows in edges outside the shortest path tree divided by the
sum of flows in the shortest path tree. Qalt can also be determined based on hydraulic
Water 2023, 15, 2896
simulation results from the resilience assessment. For the toy example in Figure 4, Qalt can
6 of 15
be determined with 0.026 (flow outside the shortest path tree 0.78 divided by the sum of
all edge flows 30.34).

Figure 3.
Figure 3. Concept
Conceptof
ofwater
waterquality
qualityassessment
assessment with
with CNA
CNA adapted
adapted from
from [15][15] (licensed
(licensed under
under CC-BY
CC-BY 4.0).
4.0).
It can be considered that the flow is not only in the shortest path tree, and corrections
basedTheon node degree describes
the network topology thecan number of edges
be applied. connected
For small WDNs, to athe
node, and mD
so-called is the
network
average number
dispersion is low, ofand edges
goodconnected
results cantoalready
a nodebeindetermined
the network. Hwang
with C0 = 0.and Lanseywith
However, [21]
evaluated 26 WDNs and determined an average mD of 2.44 and a standard
increasing network size, the network dispersion gets more important [15] and should be deviation of
±0.38. Giustolisi et al. [22] evaluated another set of 22 WDNs and determined
calibrated with simulation results of water quality of the investigated network. As global an average
mD of 2.27
network and a standard
correction, deviation
the fraction of in
of flow ±0.12. For the (alternative
the co-tree toy exampleflow,in Figure
0 ≤ Q3,altobtaining
≤ 1) and
mDmean
the = 2 = (1 + 3 +degree
node 2 + 2 +(mD)
2 + 2 +of2the
+ 2)/8.
WDN were identified to describe the global network
As local
dispersion network
[15]. correction,
Qalt is the the flow
sum of flows lengths
in edges from the source
outside shortesttopath
eachtree
demand
divided node
by
were
the determined
sum of flows in (denoted mixL)
the shortest andtree.
path used. In can
Qalt the toy
alsoexample in Figure
be determined 3, mixL
based for Node
on hydraulic
1 equals to results
simulation 0.180 km. from For
theanresilience
in-depthassessment.
understanding, thetoy
For the number of mixing
example in Figurenodes
4, Qin
alt this
can
flow
be paths were
determined withalso investigated
0.026 (flow outside(denoted number
the shortest mixing
path nodes)
tree 0.78 withbyCNA.,
divided the sume.g.,
of for
all
the toy
edge example
flows 30.34).in Figure 3, the number of mixing nodes along the shortest path (minimal
The node
residence time)degree describes
in the flow path the number
to Node of edges
2 was 4. connected to a node, and mD is the
average number of edges connected to a node in the network. Hwang and Lansey [21]
evaluated 26 WDNs and determined an average mD of 2.44 and a standard deviation of
±0.38. Giustolisi et al. [22] evaluated another set of 22 WDNs and determined an average
mD of 2.27 and a standard deviation of ±0.12. For the toy example in Figure 3, obtaining
mD = 2 = (1 + 3 + 2 + 2 + 2 + 2 + 2 + 2)/8.
As local network correction, the flow lengths from the source to each demand node
were determined (denoted mixL) and used. In the toy example in Figure 3, mixL for Node 1
equals to 0.180 km. For an in-depth understanding, the number of mixing nodes in this
flow paths were also investigated (denoted number mixing nodes) with CNA., e.g., for
the toy example in Figure 3, the number of mixing nodes along the shortest path (minimal
residence time) in the flow path to Node 2 was 4.
Water 2023, 15,
Water 2023, 15, 2896
x FOR PEER REVIEW 77of
of 15
15
Water 2023, 15, x FOR PEER REVIEW 7 of 15

Figure 4. Toy example of water quality assessment with CNA.

Figure4.4. Toy
Figure Toyexample
exampleof
ofwater
waterquality
qualityassessment
assessmentwith
withCNA.
CNA.
Based
Based on these
thesedifferent
differentcorrection
correctionterms,
terms,from thethe
shortest path values of σof
i,j the
σi,j wa-
Based onon these different correction terms, from
from shortest
the shortest path
path values
values of σi,j the the
wa-
ter age
water based
age on
based CNA
on CNAcan be
candetermined
be (A
determined CNA ).The
(A CNA
entire
).The procedure,
entire systematic
procedure, testing
systematic
ter age based on CNA can be determined (ACNA).The entire procedure, systematic testing
with hundreds
testing of different
with hundreds designsdesigns
of different and network topologies,
and network and validation
topologies, of the graph-
and validation of the
with hundreds of different designs and network topologies, and validation of the graph-
based water
graph-based quality model can be found in [15]. For the toy example with fV = 0.5, C0.5,
p=
based water quality model can be found in [15]. For the toy example with fV = 0.5,= C
water quality model can be found in [15]. For the toy example with f V p=
constant = 1 and
Cp = constant = 1Cand
0 = 0,Cthe
0 =results
0, the of CNAof
results forCNA
determining water quality
for determining waterare in Figure
quality are in5,
constant = 1 and C0 = 0, the results of CNA for determining water quality are in Figure 5,
compared to an extended
Figure 5, compared period simulation
to an extended (1 h with 1(1min
period simulation water
h with quality
1 min time
water step) with
quality time
compared to an extended period simulation (1 h with 1 min water quality time step) with
Epanet2
step) with(AEpanet2
EPS). (AEPS ).
Epanet2 (AEPS).

Figure 5. Toy example of water quality assessment with CNA (ACNA) and comparison with extended
Figure5.
Figure 5. Toy
Toyexample
exampleof
ofwater
waterquality
qualityassessment
assessmentwith
withCNA
CNA(A
(A CNA) and comparison with extended
CNA ) and comparison with extended
period simulation with Epanet2 (AEPS).
period simulation with Epanet2 (AEPS).
period simulation with Epanet2 (AEPS ).
The computational efforts for assessment of water quality in a large WDN (4021 pipes
Thecomputational
The computationalefforts
effortsforforassessment
assessmentof of waterquality
qualityininaalarge
largeWDN WDN(4021
(4021pipes
pipes
and 3558 nodes) of the graph-based model waswater determined 4.2 × 10−-6-66 less than with an
and 3558
and 3558nodes)
nodes)ofofthe
thegraph-based
graph-basedmodelmodelwas wasdetermined
determined4.24.2 × ×1010 less than than with
with anan
extended period simulation with Epanet2 [15]. However, this computational efficiency
extendedperiod
extended periodsimulation
simulationwithwith Epanet2
Epanet2 [15].
[15]. However,
However, thisthis computational
computational efficiency
efficiency can
can even be further enhanced because in the graph model of the water quality has not
can even
even be further
be further enhanced
enhanced because
because in theingraph
the graph
modelmodel
of theofwater
the water
quality quality
has nothas not
been
been determined for all nodes, e.g., often only a few nodes in the WDN are decisive for
been determined
determined for allfor all nodes,
nodes, e.g., only
e.g., often oftenaonly a few nodes
few nodes in the are
in the WDN WDN are decisive
decisive for
for water
water quality problems. These decisive nodes could be determined before the optimiza-
water quality
quality problems.
problems. These decisive
These decisive nodesbecould
nodes could be determined
determined before the before the optimiza-
optimization task
tion task (with the graph-based model or as a hybrid approach with extended period sim-
tion task
(with the (with the graph-based
graph-based model ormodel or as aapproach
as a hybrid hybrid approach with extended
with extended period period sim-
simulation
ulation with Epanet2) and, subsequently in the course of optimization, the water quality
with Epanet2)
ulation and, subsequently
with Epanet2) in the course
and, subsequently in theofcourse
optimization, the water
of optimization, quality
the water of only
quality
of only these decisive nodes can be determined with the graph model.
these decisive
of only nodes can
these decisive be determined
nodes with thewith
can be determined graph model.
the graph model.
2.4.
2.4. Case Study and Optimal Design Solutions
2.4. Case
Case Study
Study and
and Optimal
Optimal Design
DesignSolutions
Solutions
As
As a
a case
case study,
study, aa large
large realcase
real casefrom
from anan Alpine
Alpine area
area waswas used.
used. DueDue to data
to data protec-
protection
As a case study, a large real case from an Alpine area was used. Due to data protec-
tion issues,
issues, the
the real real layout cannot be shown here. However, an anonymized layout is
tion issues, thelayout cannot cannot
real layout be shown be here.
shown However, an anonymized
here. However, layout is shown
an anonymized layoutinis
shown
Figure in Figure 6. In the anonymization process, the hydraulics were fully preserved
shown6.inInFigure
the anonymization process, the hydraulics
6. In the anonymization process, thewere fully preserved
hydraulics were fully(same height
preserved
(same height
differences differences and same pipe lengths), and only the visual representation was
(same height differences and same pipe lengths), and only the visual representation The
and same pipe lengths), and only the visual representation was changed. was
changed.
case study The case study
served served approximately 100,000 with
inhabitants withsource.
one single
Thesource.
changed. The case approximately 100,000 inhabitants
study served approximately one single
100,000 inhabitants with one single WDN
source.
The
model WDNhad model
3558 had 3558
nodes nodesedges,
and 4021 edges, anddecision
therefore, decisionThe
variables. The
The WDN model had and 35584021
nodes and 4021and therefore, variables.
edges, and therefore, mean node
decision variables. The
mean node degree was 2.26 and according to Hwang and Lansey [21], and the WDN was
mean node degree was 2.26 and according to Hwang and Lansey [21], and the WDN was
Water 2023, 15, 2896 8 of 15
Water 2023, 15, x FOR PEER REVIEW 8 of 15

degree was 2.26 and according to Hwang and Lansey [21], and the WDN was a distribution
a distribution branch network with a branching index of 0.71 and an average diameter of
branch network with a branching index of 0.71 and an average diameter of 162.27 mm.
162.27 mm. The Pareto-front in Figure 6b was taken from Sitzenfrei et al. [23] and was
The Pareto-front in Figure 6b was taken from Sitzenfrei et al. [23] and was determined
determined GALAXY [17] which is based on an state-of-the-art NSGA-II algorithm for
GALAXY [17] which is based on an state-of-the-art NSGA-II algorithm for two-objective
two-objective design. A population size of 100 was used, and 300,000 generations were
design. A population size of 100 was used, and 300,000 generations were simulated (in
simulated (in total, 30 Mio number function evaluations (NFE)). The computation time on
total, 30 Mio number function evaluations (NFE)). The computation time on a desktop
a desktop computer
® (Intel® Core™ i5–6500 CPU @ 3.2 GHz, Mountain View, California,
computer (Intel Core™ i5–6500 CPU @ 3.2 GHz, Mountain View, CA, USA) was 24 weeks.
United States) was 24 weeks.

Figure6.6.(a)(a)Anonymized
Figure Anonymized layout
layout of large
of large realreal
WDS;WDS; (b) Pareto-front
(b) Pareto-front of optimal
of optimal designdesign solutions
solutions from
from meta-heuristic optimization (GALAXY).
meta-heuristic optimization (GALAXY).

Forextended
For extendedperiod
periodsimulations
simulationsfor forwater
waterquality
qualityassessment
assessmentininEpanet2,
Epanet2,aawater
water
quality time step Δt = 1 min and a total simulation time of 10 days is used. Extended
quality time step ∆t = 1 min and a total simulation time of 10 days is used. Extended period period
simulationsare
simulations areperformed
performedfor forallalldesign
designsolutions
solutionsobtained
obtainedwith
withEBC
EBC andhybrid
Q Qand hybridEBC
EBCQQ
forcomparison.
for comparison.ToTo determine
determine thethe impact
impact of numerical
of numerical dispersion, ∆t is also
dispersion, Δt isvaried
also varied be-
between
tween
60 min, 60
15 min, 15 min,and
5 min, 5 min, and 1 min.
1 min.

3.3.Results
Results
First, Q Qand hybrid EBCQQdesign are compared with the
First,the
thedesign
designsolutions
solutionsfromfromEBC
EBC and hybrid EBC design are compared with the
design
designsolutions
solutionsobtained
obtainedwith withGALAXY.
GALAXY.For Forbetter
bettervisibility,
visibility,thethefigure
figureonly
onlyshows
showsthethe
costs up to values of EUR 30 M, because the maximum cost provided by EBC Q Qis around
costs up to values of EUR 30 M, because the maximum cost provided by EBC is around
EUR
EUR2525M, M,andandasasshown
shownlate lateon,
on,design
designsolutions
solutionswith
withhigher
highercosts
costshave
haveinsufficient
insufficientwaterwater
quality
quality performance. In Figure 7a, the colored dots are design solutions determinedwith
performance. In Figure 7a, the colored dots are design solutions determined with
EBC Q Q
EBCQ. .Design
Designsolutions
solutionsfrom fromEBCEBCQabove
aboveEUREUR 1010
MM (vdesign
(vdesign<<0.5
0.5m/s)
m/s) are
aredominated
dominatedby by
design
designsolutions
solutionswithwithhigher
highervelocities.
velocities.This
Thisisisdue
duetotothe
thefact
factthat
thatthe
theresilience
resilienceindicator
indicator
used
usedininthis
thiswork
work notnot only
only considers hydraulic excess
considers hydraulic excesspressure
pressurebut butalso
alsothe
theuniformity
uniformityof
ofdiameters
diameters connected to a node [19]. This basically favors solutions,
connected to a node [19]. This basically favors solutions, where the same where the same
diam-
diameters are connected to one node. With another resilience indicator
eters are connected to one node. With another resilience indicator focusing on the hydrau- focusing on the
hydraulic performance
lic performance (e.g.,
(e.g., [2]), [2]),
this thiswould
effect effect would
not be not be observed.
observed. However, However,
solutionssolutions
starting
starting from EUR 10 M exceed water quality requirements, therefore
from EUR 10 M exceed water quality requirements, therefore investigations on this investigations on this
area
area of the Pareto-front are not
of the Pareto-front are not intensified.intensified.
The design based on EBCQ produces solutions with the minimal resilience of 0.65.
Solutions below that value are more and more driven by topography and less by topology.
For obtaining the design solutions with EBCQ, for each design solution, one single hydrau-
lic simulation run was required to determine the resilience value. Therefore, in total, 55
NFE (number of function evaluations) were required for that part of the Pareto-front.
However, when comparing these results (especially the knee bend) with those obtained
with the genetic algorithm (GA) with 30 Mio NFE, excellent results were obtained.
 The hybrid EBCQ solutions produce design solutions with lower costs. However,
more NFE were required due to the iterative nature of the procedure. For the first design
solution (In = 0.6257, EUR 4.54 M), 3 additional NFE were required, while for the least-cost
design (In = 0.3808, EUR 3.5363 M), 719 NFE were required in total. Although the results
Water 2023, 15, 2896
obtained with hybrid EBCQ in general improved the results of the CNA approach, the
9 of 15
results obtained with the GA are still slightly better (but with tremendously more compu-
tational costs).

Figure 7. (a) Comparison of designs based on genetic algorithm (GA) with graph-based design
Figure 7. (a) Comparison of designs based on genetic algorithm (GA) with graph-based design
(EBCQ and hybrid EBCQ); (b) comparison (R22) of water quality assessment (AEPS and ACNA) of the
(EBCQ and hybrid EBCQ );Q(b) comparison (R ) of water quality assessment (AEPS and ACNA ) of the
graph-based design (EBCQ and hybrid EBCQQ).
graph-based design (EBC and hybrid EBC ).
The design
The accuracy of the
based ongraph-based
EBCQ produces water quality with
solutions model theforminimal
the different assessment
resilience of 0.65.
steps is shown
Solutions belowinthat
Figure
value 7b.are
Therein,
more andthemore
design solutions
driven obtained with
by topography and the lessgraph-based
by topology.
design
For (EBCQthe
obtaining anddesign
hybrid EBCQ from
solutions withFigure
EBCQ ,7a) for are
eachassessed with extended
design solution, one single period simu-
hydraulic
lation with Epanet2. The obtained water ages (A EPS) are then compared with the results
simulation run was required to determine the resilience value. Therefore, in total, 55 NFE
obtained of
(number with graph-based
function waterwere
evaluations) quality analysis
required for (A CNA) by calculating the coefficient of
that part of the Pareto-front. However,
determination (R 2). With the plain shortest path analysis (σi,j) with residence time as edge
when comparing these results (especially the knee bend) with those obtained with the
weights,algorithm
genetic useful solutions
(GA) with can30 beMio
obtained
NFE, (R
2 up to 0.71), but in median the results are less
excellent results were obtained.
convincing (median RQ2 = 0.55). However, by
The hybrid EBC solutions produce design solutions applying the pattern
with correction
lower costs. to σHowever,
i,j, the re-

sults can be significantly improved (median R 2 = 0.95, maximum R2 = 0.97). Furthermore,

more NFE were required due to the iterative nature of the procedure. For the first design
with the topology
solution correction,
(In = 0.6257, EUR 4.54 theM),
improvement
3 additional is NFE
even were
more required,
(median Rwhile 2 = 0.97, maximum
for the least-
R 2 = 0.98). Note that the investigated design solutions represent a wide range of different
cost design (In = 0.3808, EUR 3.5363 M), 719 NFE were required in total. Although the
hydraulic
results designs
obtained withwith highly
hybrid EBCvarying flow velocities,
Q in general improved and the CNA
the results of thebased CNA approach
approach, is
capable
the resultsto produce
obtainedvaluable
with theresults.
GA are still slightly better (but with tremendously more
The proposed
computational costs).graph-based water quality model has the advantage of very fast exe-
cution The accuracy of thetograph-based
time compared extended period simulations
water quality model(EPS)
for thewith Epanet2
different but withsteps
assessment less
accuracy. In general, depending on the smallest network elements
is shown in Figure 7b. Therein, the design solutions obtained with the graph-based design (pipes), the water qual-
ity time
(EBC Q andstephybrid EBCQbe
Δt should chosen
from in Epanet2.
Figure One could
7a) are assessed withnow argue,period
extended that asimulation
larger Δt could with
be used inThe
Epanet2. order to alsowater
obtained ensure a short
ages (AEPScomputation
) are then comparedtime with a potentially
with the resultsacceptable
obtained with loss
of accuracy. Therefore,
graph-based water quality for analysis
one design(ACNAsolution from the EBC
) by calculating Q designs, of
the coefficient thedetermination
impact of Δt
(R 2 ). With
sizes was investigated. The potential
the plain shortest loss (σ
path analysis ofi,jaccuracy with a larger
) with residence time asΔt,edgecanweights,
be interpreted
useful
as numerical dispersion. 2
solutions can be obtained (R up to 0.71), but in median the results are less convincing
(median R2 = 0.55).
In Figure 8a, the water ages
However, obtained the
by applying for pattern
differentcorrection
Δt (60 min, to σ15
i,j ,min, 5 min,can
the results andbe1
min) are presented.
significantly improved The(median 2 = 0.95, maximum
boxplotsRrepresent the calculated R2 =nodal
0.97). water ages based
Furthermore, with onthe
all
topology correction, the improvement is even more (median R2 = 0.97, maximum R2 = 0.98).
Note that the investigated design solutions represent a wide range of different hydraulic
designs with highly varying flow velocities, and the CNA based approach is capable to
produce valuable results.
The proposed graph-based water quality model has the advantage of very fast exe-
cution time compared to extended period simulations (EPS) with Epanet2 but with less
accuracy. In general, depending on the smallest network elements (pipes), the water quality
Water 2023, 15, 2896 10 of 15

time step ∆t should be chosen in Epanet2. One could now argue, that a larger ∆t could be
used in order to also ensure a short computation time with a potentially acceptable loss
of accuracy. Therefore, for one design solution from the EBCQ designs, the impact of ∆t
Water 2023, 15, x FOR PEER REVIEW
sizes was investigated. The potential loss of accuracy with a larger ∆t, can be interpreted as
10 of 15
numerical dispersion.
In Figure 8a, the water ages obtained for different ∆t (60 min, 15 min, 5 min, and 1 min)
are presented. The boxplots represent the calculated nodal water ages based on all demand
demand
nodes nodes
in the in theOne
system. system. Onehow
can see, can different
see, how sizes
different
of ∆tsizes
haveofanΔtimpact
have an onimpact on the
the results of
results of the water age simulations. However, the smallest nodal
the water age simulations. However, the smallest nodal values for all solutions are close values for all solutions
arezero,
to closewhile
to zero, while
there is athere is a significant
significant difference difference in the maximum
in the maximum water ages.waterTherefore,
ages. There- in
fore, in8b,
Figure Figure 8b, thewater
the nodal nodalageswater forages
the for the different
different ∆t are
sizes ofsizes of Δt are evaluated
evaluated depend-
depending on
ing number
the on the number
of mixing of mixing
nodes fromnodesthe from the source
source to the to the demand
demand nodes.nodes. In Epanet2,
In Epanet2, at eachat
each node,
node, the inflows
the inflows are flow areweighted
flow weighted
averaged. averaged.
If thereIfisthere is aage
a water water age between
between two
two specific
specific
time time
steps, thesteps, the next
next larger larger
water agewater
is used agebased on ∆t.
is used based
Theon Δt. The maximum
maximum number ofnumbermixing
nodes in the investigated WDN is 51, meaning for the furthest distant node,distant
of mixing nodes in the investigated WDN is 51, meaning for the furthest a waternode,
parcela
water
in parcelthe
σi,j with in shortest
σi,j with residence
the shortest timeresidence time passes
passes through throughwhere
51 nodes, 51 nodes, where it is
it is potentially
potentially
mixed with mixed
flows fromwith the
flows from Due
co-tree. the co-tree. Due to the
to the rounding uprounding
to the next up∆t,
tothe
thewater
next Δt,agethe
is
water age
usually is usually overestimated
overestimated with increasing with ∆t.increasing
That effectΔt. is That
clearlyeffect
shownis clearly shown
in Figure 8b, in Fig-
where
the
ure x-axis
8b, whererepresents
the x-axis therepresents
number ofthe mixing
number nodes for reaching
of mixing nodes foreach demand
reaching anddemand
each on the
y-axis,
and onwith the ∆Age
the y-axis, with(h)—which is different is
the ΔAge (h)—which between
different thebetween
currentthe AEPS (∆t > A
current 1 EPS
min)(Δtand
>1
A
min) determined with ∆t = 1 min. For each ∆t, the slope k of the linear
EPS and AEPS determined with Δt = 1 min. For each Δt, the slope k of the linear regression regression is shown.
For example,
is shown. Forwith ∆t = 1with
example, h, anΔt average
= 1 h, of an1.72 h is additionally
average added for each
of 1.72 h is additionally mixing
added fornode.
each
For the case
mixing node.study,
For the thecase
over-estimation of water age for
study, the over-estimation 1 h ofage
of water waterfor quality time quality
1 h of water step is
up
timetostep
approximately 100 h, 15 min
is up to approximately 100uph,to 1520 h, and
min up tofor205h,min
andup forto57minh, respectively. This
up to 7 h, respec-
analysis
tively. Thishighlights
analysisthe importance
highlights of using a small
the importance of using enough
a small ∆tenough
for water quality
Δt for wateranalysis
quality
with Epanet2.
analysis with Furthermore, it demonstrates
Epanet2. Furthermore, that with larger
it demonstrates that ∆t values,
with larger only
Δt partly
values, usable
only
results are obtained.
partly usable results are obtained.

Figure 8. (a) Impact of different Δt on the water age; (b) impact of number of mixing nodes on the
Figure 8. (a) Impact of different ∆t on the water age; (b) impact of number of mixing nodes on the
water age difference ΔAge to the results with Δt = 1 min.
water age difference ∆Age to the results with ∆t = 1 min.

In Table
In Table 1,1, for
for one
one design
design solution
solution from
from the
the EBC
EBCQQ design
design (see
(see also
also Figure
Figure 8a),
8a), the
the
statistical differences in nodal water age and the computation times are shown. The graph-
statistical differences in nodal water age and the computation times are shown. The graph-
based method has by far the shortest computation time, and compared to the results with
Δt = 1 min, the best accuracy of the solution (median difference −0.19 h).
Water 2023, 15, 2896 11 of 15

Water 2023, 15, x FOR PEER REVIEW 11 of 15

based method has by far the shortest computation time, and compared to the results with
∆t = 1 min, the best accuracy of the solution (median difference −0.19 h).
Table 1. Mean differences in water age with different methods for one design solution and compu-
tation time.
Table 1. Mean differences in water age with different methods for one design solution and computa-
tion time. Median Difference (h) and Computational Time for Water Qual-
Δt/Method Standard Deviation to Δt = 1 ity Assessment of One Design Solu-
Median Difference (h) and Computational Time for Water Quality
∆t/Method min tion
Standard Deviation to ∆t = 1 min Assessment of One Design Solution
60 min −46.18 h ± 17.22 h 0.09 min = 5.5 s
60 min −46.18 h ± 17.22 h 0.09 min = 5.5 s
15 min
15 min −−10.57 h ± 3.97 h
10.57 h ± 3.97 h
0.13 min = 7.9 s
0.13 min = 7.9 s
55min
min −−2.74
2.74 hh±±1.06
1.06hh 0.88
0.88 min
min == 53.0
53.0ss
11min
min -- 30.61
30.61 min == 1836.7
1836.7ss
AACNA
CNA
−−0.19
0.19 hh±±0.27
0.27hh 0.00038 min
0.00038 min == 0.023
0.023s s

As outlined
As outlined before,
before, aa further
further advantage
advantage of of the
the graph-based
graph-based approach
approach is is that
that only
only the
the
water quality of a few nodes can be determined, avoiding even more
water quality of a few nodes can be determined, avoiding even more computational efforts. computational ef-
forts. Therefore,
Therefore, the decisive
the decisive nodes for nodes for quality
water water quality
performanceperformance are determined.
are determined. This canThis
be
done
can be with
doneextended period simulation
with extended with Epanet2
period simulation or the graph-based
with Epanet2 model. The
or the graph-based two
model.
controlling nodes (CN1
The two controlling and(CN1
nodes CN2)and are determined with a threshold
CN2) are determined with a of a water age
threshold of aexceeding
water age
36 h with A36
exceeding h
CNA . In
with the
A left
CNA . Inof
theFigure
left 9,
of the
Figureflow9, paths
the flowof CN1
paths and
of CN2
CN1 from
and CN2the source
from the
are highlighted
source in red. In
are highlighted the right
in red. In theofright
Figure 9, the cumulative
of Figure distribution
9, the cumulative function
distribution of all
function
travel path lengths
of all travel is presented,
path lengths ranging
is presented, from 0from
ranging m to 07480m tom;7480
however, the twothe
m; however, controlling
two con-
nodes CN1 and CN2 have significantly less travel path length than
trolling nodes CN1 and CN2 have significantly less travel path length than the maximum the maximum path
length, namely
path length, 5150 m5150
namely andm 3010
andm,3010respectively (red markers
m, respectively in the right
(red markers of right
in the Figureof9). This
Figure
underpins the importance
9). This underpins of considering
the importance the actual
of considering theflow regime.
actual flow regime.

Figure 9.
Figure 9. Comparison
Comparison of
of water
water quality
quality of
of controlling
controlling nodes
nodes (CN).
(CN).

For the
For thecontrolling
controllingnodes
nodesCN1CN1andand CN2,
CN2, in in Figure
Figure 10 10
thethe water
water agesages obtained
obtained withwith
the
the graph-based
graph-based method
method (ACNA (A)CNA ) and
and the the extended
extended period
period simulation
simulation withwith Epanet2
Epanet2 (AEPS(A EPS)
) for
fordesign
all all design solutions
solutions are are shown.
shown. TheThe colors
colors are are according
according to the
to the obtained
obtained resilience
resilience val-
values.
ues. can
One Onesee,can that
see, with
that with decreasing
decreasing resilience
resilience valuesvalues
(from(from
blueblue
to red to marker
red marker colors),
colors), the
the maximum
maximum water
water agethe
age in incontrolling
the controlling
nodesnodes
also also decreases.
decreases. But, But,
it canitbe
can be observed
observed that
A
that
CNAAslightly underestimates
CNA slightly underestimates thethe
water ageage
water whenwhencomparing
comparing it to AA
it to . .However,
EPS
EPS However, thisthis
underestimation
underestimation could be be compensated
compensatedby, by,e.g.,
e.g.,using
usinga safety
a safety factor
factor forfor
thethe threshold
threshold for
for
AEBCA(to
EBCbe(toon
bethe
onsafe
the side).
safe side). Following
Following the suggestion
the suggestion byC[15],
by [15], 0 = 0.1Cwas
0 = 0.1
usedwas
asused
param- as
eter value for network correction. Also, the network correction/dispersion could be
 one can see that the solutions with a high resilience value from GALAXY, all have a very
large water age (>168 h) (grey dots). When using the resilience metric according to Prasad
and Park [19], the high resilient solutions are driven by the uniformity factor, which favors
similar sized pipe diameters connected to a node. This results in large capacities all over
Water 2023, 15, 2896
the WDN. Such design solutions with almost uniformly distributed high capacity cannot 12 of 15
be obtained with the graph-based design (so far, but it will be tackled in future work).
However, when looking at the graph-based design solutions with costs above EUR 15
Mio., the maximum water age is still close to 36 h while for the design solution from GAL-
parameter
AXY abovevalue
EUR 8for network
Mio., 168 h correction.
and more are Also, the network correction/dispersion could
obtained.
be increased in the graph-based water quality
Finally, when comparing the design solutions model toEBC
of the potentially
Q designimprove
with the the results,
remaining
but this would require more investigations. As a threshold for sufficient water
part of the pareto front obtained with GALAXY (black dots), very good solutions are age in ob-
the
graph-based designs in this work, 36 h based on ACNA is used.
tained with the proposed graph-based framework.

Figure
Figure 10.
10. (a)
(a) water
water age
age comparison (ACNA and
comparison (A AEPS) for different resilience values for the controlling
CNA and AEPS ) for different resilience values for the controlling
nodes CN1 and CN2 (b) results from graph-based design (colored dots) in comparison with results
nodes CN1 and CN2 (b) results from graph-based design (colored dots) in comparison with results
from GALAXY genetic algorithm (GA) with number of function evaluations NFE = 30 Mio.
from GALAXY genetic algorithm (GA) with number of function evaluations NFE = 30 Mio.

In a last step, the obtained Pareto-front for minimizing costs and maximizing resilience
from the graph-based design method is compared with results obtained with GALAXY.
For the graph-based designs, the water quality is also assessed with the graph-based
water quality model, and design solutions with a nodal water age exceeding 36 h are
identified. For the design solutions obtained with GALAXY (population size 100 and
300,000 generations result in NFE of 30 Mio.), an EPS over 10 days with a water quality
time step ∆t = 1 min is performed for the final design solutions. Note that in Sitzenfrei [15],
it was shown that when checking the water quality constraints for design solutions from
GALAXY with the graph-based water quality model, it produces almost identical results
with EPS as with Epanet2.
When we now compare the obtained design solutions from the graph-based design
procedure (colored dots) with design solutions form GALAXY (black and grey dots) in
Figure 10b, it can be observed that for some regions of the Pareto-front, good results can be
obtained with both methods. However, when looking at the water quality performance,
one can see that the solutions with a high resilience value from GALAXY, all have a very
large water age (>168 h) (grey dots). When using the resilience metric according to Prasad
and Park [19], the high resilient solutions are driven by the uniformity factor, which favors
similar sized pipe diameters connected to a node. This results in large capacities all over
the WDN. Such design solutions with almost uniformly distributed high capacity cannot
be obtained with the graph-based design (so far, but it will be tackled in future work).
However, when looking at the graph-based design solutions with costs above EUR 15 Mio.,
the maximum water age is still close to 36 h while for the design solution from GALAXY
above EUR 8 Mio., 168 h and more are obtained.
Water 2023, 15, 2896 13 of 15

Finally, when comparing the design solutions of the EBCQ design with the remaining
part of the pareto front obtained with GALAXY (black dots), very good solutions are
obtained with the proposed graph-based framework.
The investigated WDN is a distribution branch network [21] with a common structure
for all-pipe models. For more grid like structures, the proposed approach is applicable
for determining low-cost solutions (i.e., below the knee bend in the Pareto-front) as these
design solutions are mainly driven by shortest paths, and the water age is also mainly
driven by this main distribution paths. For design solutions with higher resilience values,
there is expected to be a higher discrepancy with design solutions from meta-heuristics and
also in terms of accuracy of the graph water quality model. However, more investigations
are required in this regard.
The proposed method was shown for a single-source-fed WDN. However, multi-
source WDNs can be decomposed with graph theory as well. Sitzenfrei, Wang, Kapelan,
and Savic [23] proposed a decomposition method based on shortest path analysis. Therein,
from all sources, the shortest paths to all demand nodes were determined. With an assumed
maximum friction loss in a pipe (e.g., 1 m/km), the head losses if that node would be
supplied from each specific source were determined. These head losses for each node are
then subtracted from the water levels of the corresponding source (e.g., with four sources,
four possible supply pressures can be determined). The supply from the source with the
highest possible supply pressure is then assigned to that node, i.e., the WDN is decomposed
to supply areas. In Hajibabaei et al. [24] the sensitivity of the assumed maximum friction
losses was determined for different case studies, and for 1 m/km, the best results were
obtained also for the high resilient solutions.

4. Summary and Conclusions

This paper introduced a novel graph framework for optimizing water distribution
networks (WDNs) by combining for WDN customized approaches based on complex
network analysis. The first approach, presented in [10], is a two-objective optimization
model that was enhanced using a hybrid method for considering topography in graph
designs [12]. The second approach, described in [11], is a graph-based water quality model
which is capable of assessing water quality with significantly less computational effort than
traditional approaches. The main objective of this framework is to find Pareto-optimal
design solutions for WDNs, particularly for those that are complex and large in scale. These
solutions strike a balance between minimizing costs and maximizing resilience, with a
specific focus on controlling water age below a threshold. By considering this threshold,
technically unsuitable solutions are filtered out from the Pareto-front, ensuring that only
viable options are considered. The entire procedure is outlined in a toy example, and the
applicability and effectiveness of this novel framework is shown with a real large case study.
Remarkably, the proposed method delivers results within a few seconds of computation
time, highlighting its practicality and computational efficiency.
With the proposed graph-based methods, very fast approximate results for multi-
objective optimization can be obtained. The fast execution times enables an implementation
in, e.g., hydraulic simulation software, where a fast and user-friendly pre-assessment is
needed or any kind of application where many scenarios need to be investigated (e.g., deep
uncertainty analysis, future developments, etc.). Further, the graph-design procedure can
serve as a global search method for multi-objective optimization, which is then further
improved by a local search algorithm based on an evolutionary algorithm.
The graph-based water quality model also enables other further applications. For
example, in the course of multi-objective optimization (e.g., with GALAXY), an additional
constraint could be implemented excluding design solutions, which are (by far) exceeding
a water age threshold. By that, the search space is also reduced, potentially improving also
the computation time of the evolutionary algorithm.
In summary, this paper presents a novel graph framework that integrates complex
network analysis approaches to optimize WDNs. The results obtained through real case
Water 2023, 15, 2896 14 of 15

studies illustrate the applicability and efficiency of this approach in addressing the chal-
lenges associated with complex, large-scale WDNs.

Funding: This research was conducted within the project “RESIST”, funded by the Austrian security
research program KIRAS of the Federal Ministry of Finance (BMF). Among others, the project RESIST
aims to increase the resilience of water distribution networks by further developing graph-based
approaches for water distribution networks to practical applications.
Data Availability Statement: The anonymized case study data are freely available via https://www.
uibk.ac.at/umwelttechnik/softwareanddatasets/index.html.de (accessed on 8 August 2023). Matlab
codes are available upon request.
Conflicts of Interest: The author declares no conflict of interest.

References
1. Möderl, M.; Sitzenfrei, R.; Rauch, W. How Many Network Sources Are Enough? In Proceedings of the World Environmental &
Water Resources Congress, Challenges of Change, Providence, RI, USA, 16–20 May 2010.
2. Todini, E. Looped water distribution networks design using a resilience index based heuristic approach. Urban Water 2000, 2,
115–122. [CrossRef]
3. Farmani, R.; Walters, G.; Savic, D. Evolutionary multi-objective optimization of the design and operation of water distribution
network: Total cost vs. Reliability vs. Water quality. J. Hydroinformatics 2006, 8, 165–179. [CrossRef]
4. Zischg, J.; Klinkhamer, C.; Zhan, X.; Rao, P.S.C.; Sitzenfrei, R. A century of topological coevolution of complex infrastructure
networks in an alpine city. Complexity 2019, 2019, 16. [CrossRef]
5. Maier, H.R.; Kapelan, Z.; Kasprzyk, J.; Kollat, J.; Matott, L.S.; Cunha, M.C.; Dandy, G.C.; Gibbs, M.S.; Keedwell, E.; Marchi,
A.; et al. Evolutionary algorithms and other metaheuristics in water resources: Current status, research challenges and future
directions. Environ. Model. Softw. 2014, 62, 271–299. [CrossRef]
6. De Corte, A.; Sörensen, K. Optimisation of gravity-fed water distribution network design: A critical review. Eur. J. Oper. Res.
2013, 228, 1–10. [CrossRef]
7. Mala-Jetmarova, H.; Sultanova, N.; Savic, D. Lost in optimisation of water distribution systems? A literature review of system
operation. Environ. Model. Softw. 2017, 93, 209–254. [CrossRef]
8. Möderl, M.; Sitzenfrei, R.; Fetz, T.; Fleischhacker, E.; Rauch, W. Systematic generation of virtual networks for water supply. Water
Resour. Res. 2011, 47, W02502. [CrossRef]
9. Simone, A.; Ciliberti, F.G.; Laucelli, D.B.; Berardi, L.; Giustolisi, O. Edge betweenness for water distribution networks domain
analysis. J. Hydroinformatics 2020, 22, 121–131. [CrossRef]
10. Giudicianni, C.; Herrera, M.; di Nardo, A.; Adeyeye, K. Automatic multiscale approach for water networks partitioning into
dynamic district metered areas. Water Resour. Manag. 2020, 34, 835–848. [CrossRef]
11. Ulusoy, A.-J.; Stoianov, I.; Chazerain, A. Hydraulically informed graph theoretic measure of link criticality for the resilience
analysis of water distribution networks. Appl. Netw. Sci. 2018, 3, 31. [CrossRef] [PubMed]
12. Meng, F.; Fu, G.; Farmani, R.; Sweetapple, C.; Butler, D. Topological attributes of network resilience: A study in water distribution
systems. Water Res. 2018, 143, 376–386. [CrossRef] [PubMed]
13. Giudicianni, C.; Di Nardo, A.; Di Natale, M.; Greco, R.; Santonastaso, G.; Scala, A. Topological taxonomy of water distribution
networks. Water 2018, 10, 444. [CrossRef]
14. Sitzenfrei, R.; Wang, Q.; Kapelan, Z.; Savić, D. Using complex network analysis for optimization of water distribution networks.
Water Resour. Res. 2020, 56, e2020WR027929. [CrossRef] [PubMed]
15. Sitzenfrei, R. Using complex network analysis for water quality assessment in large water distribution systems. Water Res. 2021,
201, 117359. [CrossRef] [PubMed]
16. Sitzenfrei, R.; Qiu, M.; Ostfeld, A.; Savic, D.; Kapelan, Z. A Hybrid Approach for Considering Topography in Graph-Based
Optimization of Water Distribution Networks. In Proceedings of the World Environmental and Water Resources Congress 2023,
Henderson, NV, USA, 21–24 May 2023; pp. 831–841.
17. Wang, Q.; Savić, D.A.; Kapelan, Z. Galaxy: A new hybrid moea for the optimal design of water distribution systems. Water Resour.
Res. 2017, 53, 1997–2015. [CrossRef]
18. Sitzenfrei, R.; Oberascher, M.; Zischg, J. Identification of network patterns in optimal water distribution systems based on complex
network analysis. In Proceedings of the World Environmental and Water Resources Congress 2019, Pittsburgh, PA, USA, 19–23
May 2019; pp. 473–483.
19. Prasad, T.D.; Park, N.-S. Multiobjective genetic algorithms for design of water distribution networks. J. Water Resour. Plan. Manag.
2004, 130, 73–82. [CrossRef]
20. Rossman, L.A. Epanet 2 User Manual; National Risk Management Research Laboratory—U.S. Environmental Protection Agency:
Cincinnati, Ohio, 2000.
21. Hwang, H.; Lansey, K. Water distribution system classification using system characteristics and graph-theory metrics. J. Water
Resour. Plan. Manag. 2017, 143, 04017071. [CrossRef]
Water 2023, 15, 2896 15 of 15

22. Giustolisi, O.; Simone, A.; Ridolfi, L. Network structure classification and features of water distribution systems. Water Resour.
Res. 2017, 53, 3407–3423. [CrossRef]
23. Sitzenfrei, R.; Wang, Q.; Kapelan, Z.; Savic, D. A complex network approach for pareto-optimal design of water distribution
networks. In Proceedings of the World Environmental and Water Resources Congress 2021, Online, 7–11 June 2021.
24. Hajibabaei, M.; Hesarkazzazi, S.; Minaei, A.; Savić, D.; Sitzenfrei, R. Pareto-optimal design of water distribution networks: An
improved graph theory-based approach. J. Hydroinformatics 2023. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.